Predicting Potential Real-Time Donations in YouTube Live Streaming Services via Continuous-Time Dynamic Graph

Abstract

AbstractOnline live streaming services (e.g., YouTube Live, Twitch) are booming in recent years and gaining popularity in people’s cyber life. Real-time gifts paid by viewers in live streaming bring considerable profits and fame to streamers, whereas only a few works are interested in the donation system on live streaming platforms. In this paper, we focus on the real-time donation ‘Superchat’ on YouTube live platform and build a continuous-time dynamic graph to model the interactions among viewers based on real-time chat messages. Live streaming viewers tend to respond to the superchat immediately, demonstrating the possibility of predicting the real-time donations by analyzing other active viewers and chat messages. We design a temporal graph neural network architecture to dynamically predict the potential viewers who send donations during live streaming. Also, our model can predict the exact periods when superchat appears. Extensive experiments on three live streaming video datasets show our proposed model’s effectiveness and robustness compared to baseline methods from other fields.KeywordsOnline live streamingReal-time donationContinuous-time dynamic graphDynamic node label prediction